Consumption Inequality: Evidence and Measurement Problems

Transcription

1 : Evidence and Measurement Problems by James J. Heckman University of Chicago AEA Continuing Education Program ASSA Course: Microeconomics of Life Course Inequality San Francisco, CA, January 5-7, 2016

2 Has the increase in income inequality been matched by an equally large increase in consumption inequality? Source: Attanasio and Davis (1996). From here we might conclude that consumption inequality follows income inequality quite closely. But the issue is controversial.

3 Has the increase in income inequality been matched by an equally large increase in consumption inequality? Alternative evidence Source: Krueger and Perri (2006). Notes: LEA is after-tax labour earnings plus transfers, i.e., the sum of wages and salaries of all household members, plus a xed fraction of self-employment farm and non-farm income, minus reported federal, state, and local taxes (net of refunds) and SS contributions, plus government transfers (unemployment insurance, food stamps, and welfare). Household's consumption ND+ is the sum of expenditures on nondurables, services, and small durables (such as household equipment), plus imputed services from housing and vehicles. Each expenditure component is deated by expenditure-specic, quarter-specic consumer price indexes (CPIs).

4 Has the increase in income inequality been matched by an equally large increase in consumption inequality? Source: Krueger and Perri (2006). Notes: LEA is after-tax labour earnings plus transfers, i.e., the sum of wages and salaries of all household members, plus a xed fraction of self-employment farm and nonfarm income, minus reported federal, state, and local taxes (net of refunds) and SS contributions, plus government transfers (unemployment insurance, food stamps, and welfare). Household's consumption ND+ is the sum of expenditures on nondurables, services, and small durables (such as household equipment), plus imputed services from housing and vehicles. Each expenditure component is deated by expenditure-specic, quarter-specic consumer price indexes (CPIs). From here we may conclude that consumption inequality increased much less than income inequality. Why do we have this contradictory evidence?

5 Potential causes for contradictory evidence Datasets: Consumer Expenditure Survey, Panel Study of Income Dynamics, Nielsen Homescan Data Measure of consumption: expenditures with food, non-durables, total consumption (treatment of consumer durables?) Measures of income: wages, earnings, income, including or excluding taxes and transfers Measure of inequality: 90th-10th percentile dierence, variance, gini-coecient Unit under analysis: Individual, household, per capita, household equivalized

6 Potential causes for contradictory evidence Overall, one of the main problems is that there are no good datasets with consumption and income levels by household. At the same time, it is important to address the role of the measurement error in income and consumption comparisons.

7 Measurement error Sources: Bias in CPI (CPS-U vs. CPS-U-RS) Nonresponse for both income and consumption Under-reporting of income and some consumption items, treatment of top-coding, mean-reversion

8 Bias in the Price Deator

9 Sources of bias in the price deator (a) Quality bias Inadequate adjustments for the quality improvements in products over time. (b) Substitution bias A xed market basket does not account for the fact that people substitute away from high relative price items. (c) Outlet bias Movement of purchases toward low-price discount or big-box stores like Walmart. (d) New product bias Omission or long delay in the incorporation of new products into the CPI basket.

10 Do price deators rise unequally? Why is it important to look at measurement error of price indices when studying inequality? Using CPI or PCE implicit price deator, may yield a distorted picture of changes in living standards for groups whose consumption baskets are dierent from the aggregate bundles used in the deators. An increase in the aggregate price level associated with an increase in the price of necessities relative to luxuries, for example, will increase the level of true poverty compared with the same aggregate price increase associated with a relative increase in the price of luxuries.

11 Do price deators rise unequally? Should the price adjustment vary by income level? The standard consumption inequality literature uses a single price index. Nevertheless, there is some evidence that points to slower price increases for the bottom of the income distribution. We will develop this further.

12 How dierent are the most used deators? Often measures of income/consumption per person/household are deated using the CPI-U and CPI-U-RS price index (U stands for urban, RS stands for research series). CPI-U is inconsistent to study long time periods. CPI-U methodology changes over time, while CPI-U-RS presents an estimate of the CPI for all Urban Consumers from 1978 to present that incorporates most of the improvements made over that time span into the entire series. For example, ination rate if the current rental equivalence method of measuring the cost of homeownership had been in place prior to 1983.

13 How dierent are the most used deators? CPI-U-RS uses estimates based on BLS research covering a short period of time and extrapolated to a longer period. There have been several improvements in the CPI not incorporated into the CPI-U-RS. This is the case when improvements do not represent changes in methodology, have negligible impacts on the CPI's growth rate, or it was impossible to systematically estimate the impacts of the new methods in past years.

14 How dierent are the most used deators? The CPI-U-RS increases over at an annual rate of 3.50 percent, 0.25 points slower than the conventional CPI-U. The dierence vis-à-vis the PCE and GDP are even larger: Deator CPI-U-RS PCE GDP Recreated from Gordon (2008).

15 Literature on the deator bias Boskin commission (Boskin et al. 1996) Group of economists appointed by the Senate Finance Commission concluded that the annual bias in the CPI-U was percentage points per year in the 1980s and 1990s.

16 Literature on the deator bias Hausman and Leibtag (2003) Shows that the arrival of a Walmart store in a community reduces consumer prices for food by 25 percent, of which 20 percent is the direct Walmart impact and the other 5 percent represents the response of local stores to the Walmart arrival. What should be Walmart impact on inequality? There is no quantitative evidence on the magnitude of this eect on inequality. But notice that both because low-income households shop at Walmart, and because they spend a larger proportion of their household income on food than high-income households, there should be signicant reduction in their cost of living. The CPI ignores this eect.

17 Literature on the deator bias Hobijn and Lagakos (2005) Study dierences across households to see if there are systematic relationships between ination rates and household characteristics. No eect is found. There may be some problems with their approach: depends on ocial CPI data, which is subject to the outlet substitution bias (e.g., Walmart prices as a measured decline in the price level), focuses on deviations in household ination rates over three-month intervals rather than the long-run (more suited for the study of consumption inequality).

18 Gordon (2006) Recent changes in the methodology of CPS-U and CPS-U-RS continue to present a bias of 0.8 percentage points per year. Meyer and Sullivan (2013) Construct income series correcting for biases in price deator: substitution bias, outlet bias, quality bias and new-product bias. Concludes that the increase in inequality and poverty rates may be over estimated.

19 Meyer and Sullivan (2013) The impact of the deator can be very large Ocial and alternative income poverty rates Notes: Data are from the CPS-ASEC/ADF and CEX. Ocial Income Poverty follows the U.S. Census denition of income poverty using ocial thresholds. For measures other than the ocial one, the threshold in 1980 is equal to the value that yields a poverty rate equal to the ocial poverty rate in 1980 (13.0 percent). The thresholds in 1980 are then adjusted over time using the adjusted CPI-U-RS, which subtracts 1.1 percentage points from the CPI-U-RS each year from and 0.8 percentage points from the CPI-U-RS each year from Poverty status is determined at the family level and then person weighted. After-Tax Money Income includes taxes and credits (calculated using TAXSIM). CE data are not available for the year , ,

20 Meyer and Sullivan (2013)

21 Table 1: Real Hourly Wages at the 10th Percentile Using Alternative Deators, Source: Feenstra et al. (2002) and authors' calculations (Broda and Weinstein, 2008).

22 Evidence on Nonresponse and Under-reporting

23 Nonresponse is increasing over time in most surveys Non-response rate for wages (PSID) Recreated from PSID-Technical Series Paper #11-02.

24 Nonresponse patterns change with modications in the questions Non-response rate for spending on food consumed at home (PSID) Recreated from PSID-Technical Series Paper #11-02.

25 Who are the nonreporters? Lillard, Smith and Welch (1986) compare earnings (self-reported and imputed) and found that relation between earnings and the probability of not reporting is U-shaped-hitting its trough between $16,000 and $19,000. Proportion of Nonresponse white men by earnings interval (CPS, 1980) Income interval Earnings Earnings only Any other items 1,000-2, ,000-5, ,000-11, ,000-24, ,000-34, ,000-39, ,000-49, Over 50, Reproduced from Lillard, Smith and Welch (1986).

26 Who are the nonreporters? Non-response rates to average wage calculation from Census, 5% sample Women white black Hispanic Asian Native American Men white black Hispanic Asian Native American Source: IPUMS 5% sample.

27 Imputation methods can create bias Male and female wage-age proles (left and right, respectively) Reproduced from Bollinger and Hirch (2006). Estimates are from a pooled wage equation of respondents and imputed earners using the CPS-ORG for The male sample size s 388,578 (276,909 respondents and 111,669 imputed). The female sample size is 369,762 (270,537 respondents and 99,225 imputed). The sample includes all nonstudent wage and salary workers, ages 18 and over. Shown are log wage dierentials at each age relative to earnings of respondents who are age 18. In addition to the education dummies, control variables include race-ethnicity (four dummy variables for ve categories), foreign born, labor market size, region, and year.

28 Over- and Under-reporting Bollinger (1998) utilizes an exact match le between the 1978 March CPS and administrative records from the SSA to analyze errors in the reporting of annual income using nonparametric methodology. Their nding point to higher measurement error in cross-sectional samples than in panels and a negative relationship between measurement error. This last results is driven largely by over-reporting among male low earners.

29 Over- and Under-reporting Reproduced from Bollinger (1998). Sample of males. Income in ten thousands of dollars. The continuous line represents the conditional mean, the dashed lines represent the 95% condence bounds.

30 Literature For a general literature review on income measurement error: Bound, Brown, and Mathiowitz (2001) Surveys several studies and concludes that measurement error is mean reverting in several datasets (e.g., CPS, PSID), in the sense that persons with low earnings tend to overstate their earnings and persons with high earnings tend to understate their earnings. How are the inequality measures aected by measurement error? Gottschalk and Huynh (2010) Shows how non-classical measurement error aects some summary measures of inequality and mobility.

31 Main Datasets with Data on Expenditures

32 Datasets Consumer Expenditure Survey (CEX) Main characteristics In the past, the CEX had been conducted approximately every ten years, starting in the early 1900s and ending in Since 1980, it has been conducted on a continuous basis (repeated cross-sections). Main purpose of the survey is to collect information to be used in computing the weights for the Consumer Price Index

33 Datasets It is made of two separate and independent samples: Interview Survey Data Survey Comprehensive and detailed information about consumption expenditure and its components The categories are almost exhaustive of total consumption, with the exception of personal care items Also collects some income and demographic data

34 Consumer Expenditure Survey Main problems: CEX does a poor job at reproducing the level of expenditure in the NIPA, and it is getting worse over time (Gerner and Maki, 2004). Nondurable expenditure in 2000 dollars. Reproduced from (Attanasio Battistin and Ichimura, 2007).

35 Ratio of CEX to PEC for comparable categories Year All Durable Nondurable Owned housing Other services Reproduced from Garner, McClelland and Passero (2009).

36 Interview data and Diary data have dierent implications in terms of evolution of consumption inequality over time (Attanasio, Battistin and Ichimura, 2007; Bee, Meyer and Sullivan, 2013). Standard Deviation of log per capita monthly expenditure. Reproduced from (Attanasio Battistin and Ichimura, 2007).

37 Comparisons of CEX Diary and CEX interview to PCE Aggregates Food at Home Clothing and Shoes Reproduced from Bee, Meyer and Sullivan (2013).

38 Fraction of consumer units with zero spending by some spending categories PCE category DS IS DS-IS DS IS DS-IS DS IS DS-IS Household appliances Food Clothing materials Rent and utilities Child care Mean dierence Median dierence For more categories see Table 6 in Bee, Meyer and Sullivan (2013).

39 Higher income households are increasingly likely to underreport their expenditures relative to lower income (Aguiar and Bils, 2015). VOL. 105 NO. 9 AGUIAR AND BILS: CONSUMPTION INEQUALITY Change in Relative Income-Specic Measurement Error Table 4 Change in Relative Income-Specific Measurement Error (1) (2) (3) (4) (5) Relative mis-measurement, High-income low-income (0.06) (0.05) (0.07) (0.05) (0.05) Change, / (0.06) (0.05) (0.06) (0.06) (0.06) Change, / (0.07) (0.05) (0.06) (0.06) (0.05) Change, / (0.06) (0.05) (0.06) (0.04) (0.05) Categories included All All All Without Without durables tobacco Specification OLS WLS WLS WLS WLS First-stage instrument Income Income Lagged Income Income expenditure Notes: This table reports the change in the estimated income-specific measurement error for highest-income respondents relative to lowest-income respondents: ϕ 5 ϕ 1 from equation (6). The specification for each column is the same as in Table 3. The first row is the level for the period , and the next three rows report the change over the indicated period. Standard errors are calculated using a bootstrap with 100 replications Our Sara methodology Moreira yields a corrected measure of consumption Consumption inequality, Inequality but does

40 On top of the problems capturing expenditure levels, there is also evidence that CEX does not capture as much income as other surveys. For example, the CEX aggregate income is in average only 94 per cent of CPS aggregate income (Passero, 2009), whereas the CPS also under-reports based on NIPA. This may be because high-income CEX households are less likely to report their income accuratly and/or the very top of income distribution are under-represented in the CEX.

41 Sabelhaus et al. (2014) develops a new approach to disentangle between those two explanations. They link the average Adjusted Gross Income (AGI) by zip-code to the CEX sampled households (both respondents and non-respondents). The gure shows that households in the top AGI percentile zip-codes are 10 percent less likely to participate than the rest of the sample.

42 Moreover, households within the top AGI percentiles that do participate are more likely to have lower incomes than the households in that zip-code who did not participate. The main conclusion of Sabelhaus et al. (2014) is that under-reporting in high households is likely to explain why CEX does not capture as much income as other surveys.

43 Other datasets Panel Study of Income Dynamics Main characteristics: Longitudinal survey Starts in late 1960s Very good income and demographics Also collects information on consumption Until 1997, only food expenditure is available. Housing/utilities (most of the time) After 1997, broader measures are collected, covering now 70% of total CEX spending. Main problems: Only limited coverage of spending categories (until 1997 only food, after 1997, around 70% of total expenditures) Growth in non-response

44 Evidence on

45 Summary of contributions by the literature

46 Summary of contributions by the literature

47 Summary of contributions by the literature

48 Attanasio, Hurst and Pistaferri (2015)

49 Attanasio, Hurst and Pistaferri (2015) Contributions Use multiple datasets (in particular the CEX Diary survey, Interview data and the PSID) to check consistency of estimates Correct for the non-classical measurement error (potentially correlated to income, expenditure categories and demographics characteristics) in CE data Focus on consumption categories that were documented as well measured (Meyer and Sullivan 2012, Gerner and Maki 2004) Take a stand on the nature of measurement error Crucially, use a log linear demand system to impute total consumption using PSID

50 Attanasio, Hurst and Pistaferri (2015) Samples CEX sample Contains households of urban areas, whose respondents are aged between 25 and 65. Households that did not answer all 4 surveys are excluded. [Notice that this may create attrition bias.] PSID sample Contains households whose respondents are aged between 25 and 65. Exclude the Latino sub-sample and keep the SEO subsample. Exclude observations with outlier records in income and food consumption.

51 Attanasio, Hurst and Pistaferri (2015): CEX sample Let C it be the total consumption of household i in period t C it = K k=1 qk it, where qk it is the spending in category k (k = 1,...K ) Suppose qit 1 and q2 it are two commodities that are known to be measured without systematic error. [It is not clear why this is the case. The authors did not test for alternative pairs of goods] [Assumption 1] q 1 it a luxury and q2 it a necessity can be expressed in terms of Engel curves [Assumption 2] qit 1 = C α1 it uit 1v t 1, α 1 > 1 qit 2 = C α2 it uit 2v t 2, α 2 < 1 where α 1 and α 2 are income elasticities, vt 1 and vt 2 are aggregate factors (e.g., relative prices), and uit 1 and u2 it are unobserved idiosyncratic taste shocks.

52 Attanasio, Hurst and Pistaferri (2015) Taking the logs of the ratio between q 1 it and q2 it yields log( q1 it q 2 it )) = (α 1 α 2 ) log(c it ) + log( v 1 t v 2 t ) + log( u1 it ) uit 2 Then the cross-sectional variance is given ( ) ( Var log( q1 it qit 2)) = (α 1 α 2 ) 2 Var (log(c it )) + Var log( u1 it ( ) +2(α 1 α 2 )Cov log(c it ), log( u1 it ) uit 2 By assuming that the idiosyncratic taste shocks are uncorrelated [Assumption 3], simplies to ( ) ( ) Var log( q1 it qit 2)) = (α 1 α 2 ) 2 Var (log(c it )) + Var log( u1 it ) uit 2 u 2 it ) )

53 Attanasio, Hurst and Pistaferri (2015) We can get the variance of total consumption as [ Var 1 Var (log(c it )) = (α 1 α 2 ) 2 ( log( q1 it q 2 it )) ) ( Var log( u1 it ) uit 2 If we are willing to assume that the variance of taste shocks is invariant over time [Assumption 4],then changes (for example, between time t and time j) in the variance of log consumption can be computed as: [ ( 1 Var (log(c it )) Var (log(c ij )) = (α 1 α 2 ) 2 where the proportionality factor Var ( ) log( q1 it qit 2)) Var 1 (α 1 α 2) 2 is taken from the literature. )] log( q1 ij q 2 ij )) )]

54 Attanasio, Hurst and Pistaferri (2015) The previous procedure was applied to spending on entertainment services (q 1 it ) and expenditure on food at home (q2 it ). Why? One good has elasticity greater than 1 and the other good lower than 1 (This ensures that the proportionality factor is well dened. Nevertheless, one just need that α 1 α 2 > 0) Components relatively well measured (according to evidence on Meyer and Sullivan, 2012)

55 again are striking. Across all the other measures we consider where, to reiterate, measurement error is less of an issue consumption inequality has increased by only slightly less than the increase in income inequality. Attanasio, Hurst and Pistaferri (2015) 6.1 CE Diary Data: Total Expenditure As we mentioned in the data section, the CE Survey is made of two components: the Interview and the Observed SD of log equivalized non-durable consumption Diary Survey. While the figures we have considered so far are derived from the former, analogous figures can be constructed using the latter, especially after 1986, when the Diary Survey became comprehensive and includes virtually all consumption categories. In Figure 9, we plot the standard deviation of log total consumption, for the period. Again, we adjust the data for differences in family size. St. Dev Diary Data Year Estimated SD of log ratio of entertainment/food spending (mean) sd_leisure_foodin Diary Data Inequality Ratio Entertainment-Food Year Figure 9: Standard deviation Log equivalized nondurable consumption, CE Diary Data. Figure 11a: SD log ratio Entertainment/Food Spending: CE Diary Data According to Aguiar and Bils (2012) α 1 α 2 = 1.4. Instead, The ratio between food at home and entertainment expenditure can also be computed in the Interview When comparing Figure 6 and 9 two features emerge. First, the level of inequality measured Figure 9 is data. In Figure 11b, we report the path of the standard deviation of the log of such a ratio, to be compared considerably according larger. This is notto particularly thesurprising estimates because of the structure inoflechene the two Surveys: the diary and Levell (2012) no to Figure 11a. 24 adjustment is needed as α 1 α 2 = 1. We find that, in the case of this ratio, the inequality measure that emerges from the Interview CE data is considerably larger than what obtained with total nondurable consumption expenditure and not inconsistent with the evidence coming from Consumption the Diary survey. Inequality particular, depending on when one starts

56 Attanasio, Hurst and Pistaferri (2015): PSID sample Main problem with PSID: Until 1997, it only includes measures of food consumption (food at home, away from home, and the value of food stamps). After 1997, broader measures are collected, covering around 70% of total CEX spending: Includes spending on utilities (electricity, heating, water, miscellaneous utilities), home insurance premiums, health (health insurance premiums, nursing care, doctor visits, prescriptions, other health spending), vehicle spending (vehicle insurance premiums, vehicle repairs, gasoline, parking), transportation (bus fares, taxi fares, other transportation expenses), education (tuition, other school expenses), and child care.

57 Attanasio, Hurst and Pistaferri (2015): PSID sample To create a measure of total consumption, the authors used imputation methods: Ziliak (1998) Consumption is dened as the dierence between income and the changes in assets (sum of liquid assets and equity, where the dierence between the self-reported home value and the remaining principal on the home mortgage).

58 Attanasio, Hurst and Pistaferri (2015) Blundell, Pistaferri and Preston (2008) 1 Estimate a food demand equation using CEX data ln food CEX it = X CEX it β t + ln C CEX it γ t (Eit CEX ) + ε CEX it where X it includes number of children, a quadratic in the household head's age, a dummy for self-employment, education dummies; ln C it γ t (E it ) includes log consumption and the interaction between log consumption and education; and ε it is an idiosyncratic taste preference shifter. β CEX t 2 Using and with PSID data by computing CEX γ t, one can get a measure of consumption ln Cit PSID = ln food PSID it Xit PSID β t CEX γ t CEX (Eit PSID )

59 Attanasio, Hurst and Pistaferri (2015) Notes: Imputed total consumption using BPP's procedure, Imputed total consumption using Ziliak's procedure. Observed consumption categories available post Observed consumption categories available post-1997, except expenditures with health and education. Observed food consumption.

60 Attanasio, Hurst and Pistaferri (2015) Notes: (top right) Imputed using BPP's procedure, (top left) Imputed using Ziliak's procedure, (bottom) Observed food consumption.

61 Attanasio, Hurst and Pistaferri (2015) Notes: (right) Imputed using BPP's procedure, (left) Imputed using Ziliak's procedure.

62 Meyer and Sullivan (2011) Construct income series correcting for biases in price deator Account for the role of taxes and transfers Correct CE Survey consumption measures accounting for durables and misreporting

63 Meyer and Sullivan (2011) Notes: Authors calculations using CEX. All measures are reported in 2005 dollars using adjusted CPI-U-RS, are calculated at the family level, are person weighted, and are adjusted for dierences in family size using the NAS recommended equivalence scale. Each scale adjusted measure is multiplied by 2.14, the mean adult equivalent value across all years. The consumption excluded MOOP, education and retirement spending.

64 Meyer and Sullivan (2011) Median 10th Percentile Notes: Authors calculations using CEX and CPS. All measures are reported in 2005 dollars using adjusted CPI-U-RS, are calculated at the family level, are person weighted, and are adjusted for dierences in family size using the NAS recommended equivalence scale. Each scale adjusted measure is multiplied by 2.14, the mean adult equivalent value across all years. The consumption excluded MOOP, education and retirement spending. Non-cash benets include food stamps and housing and school lunch subsidies.

65 Meyer and Sullivan (2013) Consumption and Income Inequality and the Great Recession Note: Income is after-tax money income plus food stamps and housing and school lunch subsidies. Consumption is adjusted for under-reporting by calculating a predicted value of consumption from a regression of unadjusted consumption on core consumption and demographic characteristics using data from 1980 and 1981.

66 Meyer and Sullivan (2013) Real Changes in income (left) and Consumption (right) at various percentiles Note: Income is after-tax money income plus food stamps and housing and school lunch subsidies. Figures are adjusted for ination using the adjusted CPI-U-RS. Consumption is adjusted for under-reporting by calculating a predicted value of consumption from a regression of unadjusted consumption on core consumption and demographic characteristics using data from 1980 and See text for more details. Figures are adjusted for ination using the adjusted CPI-U-RS.

67 Meyer and Sullivan (2013) Income Inequality Notes: All measures other than the o cial measure, are adjusted for di erences in family size using the NAS recommended equivalence scale. The unit of observation for the o cial measure is the household, while it is the family for the other income measures.

68 Meyer and Sullivan (2013) Income Inequality Notes: All measures other than the o cial measure, are adjusted for di erences in family size using the NAS recommended equivalence scale. The unit of observation for the o cial measure is the household, while it is the family for the other income measures.

69 Meyer and Sullivan (2013) Income Inequality Notes: All measures other than the o cial measure, are adjusted for di erences in family size using the NAS recommended equivalence scale. The unit of observation for the o cial measure is the household, while it is the family for the other income measures.

70 Meyer and Sullivan (2013) Notes: Consumption data are from the CE and income data are from the CPS. Consumption is measured as the predicted value of consumption from a regression of total consumption on core consumption and demographic characteristics using data from 1980 and Core Consumption includes consumption of housing, food at home, vehicles, and other transportation. See text for more details.

71 Meyer and Sullivan (2013) Notes: Consumption data are from the CE and income data are from the CPS. Consumption is measured as the predicted value of consumption from a regression of total consumption on core consumption and demographic characteristics using data from 1980 and Core Consumption includes consumption of housing, food at home, vehicles, and other transportation. See text for more details.

72 Meyer and Sullivan (2013) Notes: Consumption data are from the CE and income data are from the CPS. Consumption is measured as the predicted value of consumption from a regression of total consumption on core consumption and demographic characteristics using data from 1980 and Core Consumption includes consumption of housing, food at home, vehicles, and other transportation. See text for more details.

73 Meyer and Sullivan (2013) Decomposition of Changes in Notes: Data are from the CE survey. These estimates are for log consumption where consumption is measured as the predicted value of consumption from a regression of total consumption on core consumption and demographic characteristics using data from 1980 and See text for more details.

74 Meyer and Sullivan (2013) Decomposition of Changes in (cont.) Notes: Data are from the CE survey. These estimates are for log consumption where consumption is measured as the predicted value of consumption from a regression of total consumption on core consumption and demographic characteristics using data from 1980 and See text for more details.

75 Meyer and Sullivan (2013) Decomposition of Changes in (cont.)

76 Meyer and Sullivan (2013) Decomposition of Changes in (cont.)

77 Meyer and Sullivan (2013) Income Inequality Notes: All measures other than the o cial measure, are adjusted for di erences in family size using the NAS recommended equivalence scale. The unit of observation for the o cial measure is the household, while it is the family for the other income measures.

78 Meyer and Sullivan (2013) Income Inequality Notes: Consumption data are from the CE and income data are from the CPS. Consumption is measured as the predicted value of consumption from a regression of total consumption on core consumption and demographic characteristics using data from 1980 and Core Consumption includes consumption of housing, food at home, vehicles, and other transportation. See text for more details.

79 Meyer and Sullivan (2013) Income Inequality Notes: Consumption data are from the CE and income data are from the CPS. Consumption is measured as the predicted value of consumption from a regression of total consumption on core consumption and demographic characteristics using data from 1980 and Core Consumption includes consumption of housing, food at home, vehicles, and other transportation. See text for more details.

80 Meyer and Sullivan (2013) Income Inequality Notes: Consumption data are from the CE and income data are from the CPS. Consumption is measured as the predicted value of consumption from a regression of total consumption on core consumption and demographic characteristics using data from 1980 and Core Consumption includes consumption of housing, food at home, vehicles, and other transportation. See text for more details.

81 Measurement problems are also relevant when studying income inequality... Burkhauser, Larrimore, and Simon (2012, NBER) Burkhauser, Feng, Jenkins, and Larrimore (2012, ReStat) Armour, Burkhauser, Larrimore (2013, AER)

82 Have middle class American failed to benet from recent economic growth? Findings suggest that the middle class is not sharing proportionately in the fruits of American economic growth: Income of middle class households as measured by median household income (ination-adjusted) has consistently grown over time in most business cycles. (CPS data) There is evidence that the fraction of market income going to the top 10 percent of tax units is at its highest level since at least (IRS administrative records by Piketty and Saez, 2003, and Saez, 2009) Burkhauser and coauthors show that this evidence is far from clear.

83 A second opinion This paper uses cross-sectional data to capture the economic resources available to individuals at the same point in the distribution over time. Under dierent criteria, several income series are constructed. The evidence suggests that the middle class decline is far from clear, and that such results are highly sensitive to how available resources are measured and assumptions on the unit under analysis.

84 Main data sources What data was used in previous papers? IRS tax record data: Data on tax units (population that le taxes), available since 1917 Contains information on the pre-tax, pre-transfer cash income Aected by schemes to limit tax liabilities (more easily done by top income units)

85 Main data sources Annual March CPS: Data on households (sample of population), available since 1967 Contains pre-tax, post-transfer cash income excluding capital gains. Also includes the value of all public transfers (including welfare, Social Security, and other government provided cash assistance), much of which is not taxable. Aected by top coding, under report, undercoverage

86 Data sources Is the source of data relevant? Burkhauser and coauthors reconcile estimates from both sources. Most evidence results from March CPS data (non-public version) Supplemented with cell-means to overcome top-coding of high incomes (Larrimore et al. 2008). Adjusted for ination to 2008 dollars using the CPI-U-RS.

87 Data sources To overcome the fact that CPS does not directly inquire about tax credits, tax liabilities, or about the value of in-kind compensation (such as employer or government provided health insurance), they impute this information for each individual using NBER TaxSim 9.0. To overcome the fact that CPS does not accounts for the ex-ante value of in-kind health insurance benets, the authors impute cell means of employer contributions from the Medical Expenditure Panel Survey Insurance Component (MEPSIC).

88 Assumptions Exploring the importance of assumptions: measures of income unit of analysis: sharing units: household, families, and tax units unit of comparability: household, household equivalized

89 Measures of income Pre-tax, pre-transfer (market) income: includes income from wages and salaries, interest, dividends, rents, trusts, and retirement pension income but excludes transfers which are not included in market income (it also does not include capital gains) Pre-tax, post-transfer income: Adds cash transfers, including income from welfare transfer programs such as AFDC/TANF as well as from social insurance programs such as Social Security and Workers' Compensation. It excludes transfers directly tied to the tax system such as the Earned Income Tax Credit. It also excludes any in-kind government transfers, such as the value of Medicare or Medicaid insurance.

90 Measures of income Post-tax, post-transfer income: incorporates tax credits and liabilities. Post-tax, post-transfer income plus health insurance: includes ex-ante value of employer provided health insurance and ex-ante value of government provided health insurance via Medicaid and Medicare.

91 Unit of analysis If you are doing a study of wage rates or earnings, the natural unit is the individual. When studying people's economic wellbeing, the individual is certainly not the right unit of analysis. People live and share income. So, researchers need to think carefully to get the sharing unit correct. There are three dierent sharing units: Household (people that live together) Family (blood relative or who you are married to) Tax unit (Who you put on your income tax form when you pay your money to the government. Typically consists of an adult, his or her spouse, and any dependent children.)

92 Unit of analysis Which is the more appropriate unit for the sharing unit? Does it make a dierence when studying inequality? In a traditional family arrangement, a tax unit would be equivalent to household (and to family). However, exceptions to such traditional households are becoming more common. For example, cohabiters, roommates who share expenses, children who move back in with their parents (so-called boomerang kids) or older parents who live with their adult children will contain more than one tax unit. Burkhauser and coauthors study this and conclude that the distinction between tax units and households as sharing units is not trivial and cannot be treated interchangeably.

93 Unit of analysis To compare sharing units we need to adjust for the size of the household (often called equivalized measures). - Measures that do not account for the composition of the household, consider measures of income of the sharing unit with the following adjustment: y = Y, θ [0, 1] nθ If θ = 1, then it is income per capita and it does not account for economies of scale. If θ = 0, we impost perfect economies of scale. It is common to use θ = 1. That is, a sharing unit of four needs exactly 2 twice as much as a sharing unit of one to be at the same level of economic well-being. - Alternative measures such as the ones used by OECD take into account the composition of the sharing unit.

94 Results Table 1: Comparing the total growth from using each sharing unit, sizeadjustment, and income series combination. Tax Unit Household Size- Adjusted Tax Unit Size- Adjusted Household Pre-tax, pre-transfer 3.2% 12.5% 14.5% 20.6% Pre-tax, post-transfer 6.0% 15.2% 17.0% 23.6% Post-tax, post-transfer 9.5% 20.2% 25.0% 29.3% Post-tax, post-transfer + Health Insurance 18.2% 27.3% 33.0% 36.7% Source: Public Use March CPS data. Note: Changes in income between 1992 and 1993 are suppressed and assumed to be zero given the trend-break resulting from the CPS redesign in those years. See main text for details. 1 Health insurance information not available prior to The rate of growth in the value of health insurance from is assumed to match that of post-tax, post-transfer income.

95 Results The upper-left corner of the table matches the denitions used by Piketty and Saez (2003). Moving to broader denitions of income improves the median income. Using tax units also limits the measured income growth. Why? number of tax units per household increased increase in the number of cohabiters and adult children living with their parents Using adjustments for the size of the household also increases the median income. average number of individuals per household decreased

96 Results Table 2: Trends in the size of tax units and households. Panel A: Tax Units per Household Mean Tax Units per Household Percent of Households with one, two, or more Tax Units One Two Three Tax Units (Thousands) Households (Thousands) ,958 79, ,705 93, , , , , Unrelated Tax Units (Thousands) Panel B: Unrelated Tax Units per Household Mean Percent of Households with Unrelated Tax one, two, or more Unrelated Households Units per Tax Units (Thousands) Household One Two Three ,690 79, ,606 93, , , , , Panel C: Individuals per Tax Unit

97 (Thousands) (Thousands) Household One Two Three ,690 79, Results ,606 93, , , , , Panel C: Individuals per Tax Unit Mean Individuals per Tax Unit Percent of Tax Units with one, two, or more Individuals One Two Three Individuals (Thousands) Tax Units (Thousands) ,965 98, , , , , , , Panel D: Individuals per Household Mean Individuals per Household Percent of Households with one, two, or more Individuals One Two Three Individuals (Thousands) Households (Thousands) ,965 79, ,886 93, , , , , Source: See Table 1.

98 Results Table 3: Growth in median incomes using alternative income series Panel A: Total median income growth in each business cycle Tax unit Pre-tax Pretransfer Household Pre-tax Posttransfer Household Size-adj. Pre-tax Post-transfer Household Size-adj. Post-tax Post-transfer Household Size-adj. Post-tax Post-trans. + Health Ins % 6.6% 9.2% 12.0% 12.0% % 9.3% 13.4% 14.4% 16.6% % -1.2% -0.1% 1.0% 4.8% % 15.2% 23.6% 29.3% 36.7% 1 Panel B: Annualized median income growth in each business cycle Household Tax unit Pre-tax Pretransfer Household Pre-tax Posttransfer Household Size-adj. Pre-tax Post-transfer Household Size-adj. Post-tax Post-transfer Size-adj. Post-tax Post-trans. + Health Ins % 0.66% 0.92% 1.20% 1.20% % 0.85% 1.22% 1.31% 1.51% % -0.17% -0.02% 0.14% 0.68% % 0.54% 0.84% 1.05% 1.31% 1 Source: See Table 1.